1. No category

Automation in Construction 56 (2015) 14–25
Contents lists available at ScienceDirect
Automation in Construction
journal homepage: www.elsevier.com/locate/autcon
A query expansion method for retrieving online BIM resources based on
Industry Foundation Classes
Ge Gao a,d, Yu-Shen Liu a,b,c,⁎, Meng Wang a, Ming Gu a, Jun-Hai Yong a
a
BIM Research Group, School of Software, Tsinghua University, Beijing, China
Key Laboratory for Information System Security, Ministry of Education of China, China
Tsinghua National Laboratory for Information Science and Technology, China
d
Department of Computer Science and Technology, Tsinghua University, China
b
c
a r t i c l e
i n f o
Article history:
Received 2 July 2014
Received in revised form 1 April 2015
Accepted 14 April 2015
Available online xxxx
Keywords:
Building Information Modeling (BIM)
Information retrieval
Industry Foundation Classes (IFC)
Ontology
Query expansion
Local context analysis (LCA)
a b s t r a c t
With the rapid popularity of Building Information Modeling (BIM) technology, BIM resources such as building
product libraries are growing rapidly on the World Wide Web. As a result, this also increases the difficulty for
quickly finding useful BIM resources that are sufficiently close to user's specific needs. Keyword-based search
methods have been widely used due to their ease of use, but their search accuracy is often not satisfactory
because of the semantic ambiguity of terminologies in BIM-specific documents and queries. To address this
issue, we develop a prototype semantic search engine, named BIMSeek, for retrieving online BIM resources.
The central work consists of two parts as follows. Firstly, based on Industry Foundation Classes (IFC) which is a
major standard for BIM, a domain ontology is constructed for encoding BIM-specific knowledge into the search
engine. Using the ontology, term`inologies in BIM documents can be disambiguated and indexed. Secondly, by
combining the ontology and local context analysis technique, an automatic query expansion method is presented
for improving retrieval performance. Compared with traditional keyword-based methods and WordNet-based
query expansion methods, the experimental results demonstrate that our method outperforms them. The search
engine is available at http://cgcad.thss.tsinghua.edu.cn/liuyushen/ifcqe/.
© 2015 Elsevier B.V. All rights reserved.
1. Introduction
Building Information Modeling (BIM) technology has been receiving
an increasing attention in the AEC (Architecture, Engineering and
Construction) industry [1]. Compared with the traditional Computer
Aided Design (CAD) technology, BIM is capable of restoring both geometric and rich semantic information of building models, as well as
their relationships, to support lifecycle data sharing. With the rapid
popularity of BIM technology in the AEC field, BIM resources such as
building product libraries are growing rapidly on the World Wide
Web (WWW). For instance, the well-known Autodesk Seek [2] is an
online system, which provides a large repository of building products
on its website and allows users to find a wide variety of BIM products
from manufacturers. It currently carries more than 65,000 commercial
and residential building products from nearly 1000 manufacturers,
and is still growing daily. BIMobject [3] is another widely visited
website, which contains over 450,000 BIM models with the product
data and properties. Other online libraries such as National BIM Library
⁎ Corresponding author at: School of Software, Tsinghua University, Beijing 100084,
China. Tel.: +86 10 6279 5455; mobile: +86 159 1083 1178.
E-mail addresses: [email protected] (G. Gao), [email protected]
(Y.-S. Liu), [email protected] (M. Wang), [email protected] (M. Gu),
[email protected] (J.-H. Yong).
URL: http://cgcad.thss.tsinghua.edu.cn/liuyushen/ (Y.-S. Liu).
http://dx.doi.org/10.1016/j.autcon.2015.04.006
0926-5805/© 2015 Elsevier B.V. All rights reserved.
[4], Google 3D Warehouse [5], SmartBIM [6] and many active online
communities (e.g., RevitCity [7]) also contain a large number of
information contents of building products related to BIM models.
However, the large amount of online BIM resources also increases
the difficulty for quickly finding useful information that are sufficiently
close to user's specific needs. In the engineering field, some studies have
reported that engineers spent a large amount of time in searching for
information [8–10]. In the AEC field, the percentage of time spending
on search might be higher. To search online BIM documents quickly
and accurately, existing information retrieval approaches should be
appropriately adopted for possible improvement. The current practice
in information retrieval (IR) mostly relies on keyword-based search
methods, which can provide an easy way to quickly retrieve documents.
However, search accuracy of traditional keyword-based retrieval
models, such as Boolean model, vector space model, or probabilistic
model, has been often problematic because of the semantic ambiguity
of terminologies in BIM documents and queries. The semantic ambiguity in BIM documents can be alleviated by using a domain ontology.
Meanwhile, the user's query can be expanded with domain-specific
terms to improve the accuracy of the search result.
This paper aims to study the particular problem for retrieving
online BIM documents. To achieve this purpose, we develop a prototype semantic search engine, named BIMSeek, by combining
the usability of keyword-based interface with automatic query
G. Gao et al. / Automation in Construction 56 (2015) 14–25
expansion techniques. The central work consists of two parts as
follows. Firstly, based on Industry Foundation Classes (IFC) [11]
which is a major standard for BIM, a domain ontology named
IFC IR Ontology is constructed for encoding BIM-specific knowledge
into the search engine. Using the ontology, terminologies in BIM
documents can be disambiguated and indexed. Secondly, by combining the domain ontology and local context analysis (LCA) technique [12,13], an automatic query expansion method is presented
for improving retrieval performance. Here, LCA is used to select
query expansion terms based on co-occurrence with the query
terms within the top-ranked documents. Compared with traditional
keyword-based methods and WordNet-based query expansion
methods, the experimental results demonstrate that our method
outperforms them.
The remainder of the paper is organized as follows. Section 2 reviews
the related work. Section 3 gives the objective of our study. Section 4
introduces the construction process of IFC-based ontology. Section 5
describes our query expansion algorithm based on the ontology.
Section 6 illustrates our search engine and demonstrates the experimental results. Finally, Section 7 concludes the paper and discusses
the limitations and future work.
2. Related work
Most traditional keyword-based search approaches are based on the
vector space model (VSM) [14]. In this model, documents and queries
are represented as a vector of term weights and the retrieval is done
according to the similarity between these vectors. In essence, such
approaches try to derive the meaning of the text from the observable
syntactic and statistical behaviors without representing the meaning
directly. However, these approaches have the ambiguity problem for
retrieving online BIM documents. The main reason for the problem is
that BIM documents are different from general documents because of
their syntax variations and semantic complexities. For instance, online
BIM documents are often organized by different information providers,
and it is not easy to formulate the well-designed queries for the endusers. As a result, some documents, which do not contain the query
terms, may not be returned to the user even though they are semantically relevant to the given query.
To handle such problem, one possible way is to use semantic-based
IR approaches, in which the search does not completely rely on exact
term matching. Instead, such approaches usually seek to improve
retrieval accuracy by attempting to encode the user's searching intent
and the contextual meaning of terms in documents. Semantic-based
IR approaches could be roughly categorized as either explicit or implicit,
according to whether an explicit concept space is defined and utilized.
An implicit technique aims to analyze and explore the hidden relationships between a set of documents and the terms, without requiring the
concepts defined explicitly. Existing implicit approaches include local
context analysis (LCA) [12,13], latent semantic analysis (LSA) [15],
probabilistic LSA (pLSA) [16], latent Dirichlet allocation (LDA) [17],
etc. In contrary, an explicit technique tries to extract the explicit
semantic definitions between words and terms as well as the relationships between them, such as synonym, hyponym and hypernym.
These definitions and relationships are usually represented in some
form of linked data such as taxonomy, thesaurus or ontology. Existing
explicit approaches such as ontology-based query expansion [18],
semantic indexing [19], explicit semantic analysis (ESA) [20] and granularity analysis [21] have been developed for improving the semanticbased retrieval.
In the last few years, some semantic-based IR techniques have been
devoted to the study of information retrieval systems for the engineering field [9,10,22–27]. Many of them dealt with semantic representation
or domain ontology construction for applying to various engineering
fields. Li et al. [10] developed an engineering ontology in mechanical
design and manufacturing, and applied the ontology to index and
15
retrieve unstructured engineering documents and CAD drawings. Lin
et al. [23] introduced some text-based information retrieval experiments in the AEC industry, and described an AEC online product search
engine by incorporating domain knowledge and IR techniques [24].
Rezgui [27] used domain knowledge to formulate an ontology that assists in indexing and retrieving construction documents in e-COGNOS
system. Weissman et al. [26] proposed a computational framework for
authoring and searching product design specification documents using
semantic mapping. Demian and Balatsoukas [22] investigated the
effects of search result interfaces (particularly the aspects of granularity
and context) in systems for searching archives of construction documents. Lin et al. [25] presented a passage partitioning approach
according to domain ontology, and applied the approach to a conceptbased IR system for engineering domain-specific technical documents.
More recently, Hahm et al. [9] introduced a personalized query expansion approach for engineering document retrieval based on a selfconstructed engineering ontology.
Among the IR techniques in the engineering field, it is noteworthy
that query expansion approaches [18] which aim to overcome the
ambiguity of natural language and also the difficulty in using a single
term to represent an abstract concept. It is generally conducted by
supplementing original queried terms by morphological variations or
semantically related terms, so the performance degradation caused by
syntax variation and semantic complexity of BIM documents can be
overcome with query expansion. Query expansion can be classified
as interactive, manual, or automatic approaches according to their
strategies [18], where automatic query expansion is an alternative
strategy for domain-specific retrieval by employing taxonomy (e.g.,
WordNet) or domain-specific ontology. Considering the characteristics
of BIM documents, automatic query expansion based on BIM-specific
ontology is an appropriate strategy because ambiguous and complicated queried keywords can be disambiguated and interpreted by the
ontology. Also, terminologies in BIM documents can be disambiguated
and indexed with the ontology. As a result, the user's short queries
can be expanded and matched to syntax varied and semantically
complicated documents. By combining query expansion with domain
ontology, there have been several search engines on information
retrieval in the engineering domain [9,10,24,25,27]. However, they are
still limited to their ontology sources and the needs of particular
applications as follows. Firstly, most of existing studies focus on retrieving the controlled collection of engineering documents, rather than
Web-centric BIM resources. Secondly, existing methods are mainly
based on explicit semantic-based IR techniques, which are limited
to the capability of hand-crafted ontology or thesaurus and cannot
fully fulfill the retrieval task for dynamically changed Web-based
BIM resources. Thirdly, ontologies of existing IR approaches lack the
utilization of BIM-specific domain knowledge for online document
retrieval.
As the commonly used data exchange standard for BIM, Industry
Foundation Classes (IFC) [11] led by the buildingSMART, formerly
known as International Alliance for Interoperability (IAI), plays a crucial
role to facilitate interoperability between various software platforms in
the AEC industry. To date, the IFC standard has been widely supported
by the market-leading BIM software vendors. As the most widely used
taxonomy and specification in BIM applications, the underlying IFC
specification is therefore our preferred candidate semantic resource,
which provides a sharable skeleton on which the BIM-oriented IR
ontology can be built. Recently, several IFC-based ontologies have
been studied for particular application needs [28–32]. For instance,
Pauwels et al. [31] applied an IFC ontology to semantic rule checking.
Beetz et al. [29] presented an approach for converting the IFC schema
into the OWL format, which is a remarkable effort to lift the IFC
specification onto the ontology level. Zhang and Issa [28] used an IFC ontology to extract partial model from a complete IFC model. In addition,
several applications also used IFC ontologies for querying spatial information within a BIM model [32,33].
16
G. Gao et al. / Automation in Construction 56 (2015) 14–25
The common strategies mentioned in previous studies [28–31] are
to utilize their respective ontologies for extracting specific information
from an IFC file itself, whereas the currently available IFC files are
relatively few on the WWW. In practice, a large number of available
online BIM resources consist of BIM components/models (in their native
file formats but without providing IFC files), and these models are
associated with their online product documents (e.g., descriptions or
specifications). Therefore, merely parsing and retrieving IFC files cannot
take full advantage of the abundant online BIM resources. In contrast,
this paper develops a BIM-specific IR ontology based on the IFC schema,
and applies the ontology to retrieve online BIM documents rather than
IFC files. To alleviate the insufficiency of hand-crafted ontology, our
method combines the explicit semantic technique (i.e., ontologybased query expansion) with the implicit semantic technique (i.e.,
local context analysis [13]). Consequently, our search engine can cover
more widely online BIM resources for information retrieval.
3. Objective and methodology
Building product webpage on BIM resource libraries (e.g., [2,4,6])
typically contains BIM models associated with product documents
(e.g., specifications and descriptions of the objective products), as
shown in Fig. 1. The BIM models are normally in their native file format
dependent on various BIM software vendors (e.g., Autodesk Revit,
Bentley Architecture and Graphisoft ArchiCAD) or in industry-neutral
file format (e.g., IFC/ifcXML). The relevant product documents on the
Web are the textual contents for describing BIM models including
their functions, dimensions, materials, performances, manufacturers,
etc. These product documents are independent of BIM model file
format. In particular, much information is embedded in textual BIM
documents generated during design and construction phases [22].
Most online BIM documents are unstructured, in contrast to structured
contents (e.g., BIM models or database tables) following a strict schema.
In order to retrieve online BIM resources, we have developed a
prototype search engine, i.e., BIMSeek, which contains two individual
modules: document search module and model search module. The former
that will be described in this paper focuses on the problem of retrieving
online BIM documents by incorporating the domain ontology and query
expansion techniques. The latter that is an ongoing work deals with the
problem of searching online BIM models by incorporating the same
ontology and structured query techniques. Although the two modules
in BIMSeek process different contents (documents or models) of online
BIM resources, they are all based on the common ontology, named IFC IR
Ontology developed in this paper (see Section 4), for the semantic
search engine. In this paper, only the hierarchical and enumeration
relationships of the ontology are utilized for retrieving BIM documents.
However, in the model search module, the properties and restrictions of
the same ontology are also utilized for searching BIM models, where
BIM models are converted into OWL (Ontology Web Language)
instances and the user queries are transformed into SPARQL. Also, it is
interesting to make use of document and model information together
for improving search accuracy and performance in BIMSeek. In this
Fig. 1. Illustration of building product webpages that contain BIM models associated with product documents (including product specification and description).
G. Gao et al. / Automation in Construction 56 (2015) 14–25
way, more relevant BIM resources might be retrieved for the user, and
we leave this study to our future work. In addition, the IFC ontology
built in this paper is also used for another module, named BIMTag
[34], as part of BIMSeek. Using the ontology, BIMTag achieved semantic
annotation for online BIM documents.
The purpose of this study is to use the ontology to process the
contextual meaning of terms for retrieving online BIM documents.
Based on the ontology, we typically choose automatic query expansion
techniques to enhance retrieval performance by reflecting the user's
needs. Since automatic query expansion techniques require no effort
on the part of the user, they have a significant advantage over manual
techniques. Automatic query expansion can be categorized as either
global or local [13]. Global techniques rely on analysis of the whole
corpus for the source of expansion terms, while local techniques process
a small number of top-ranked documents retrieved for a query to
expand that query. In information retrieval, query expansion is the
first step of semantic search [18,19], which reformulates a given query
to increase the likelihood of the term overlap between the query and
documents that are likely to be relevant to the user's needs. If used
together with an effective word sense disambiguation (WSD) algorithm, query expansion can improve retrieval performance. However,
simply using a thesaurus or ontology for automatic query expansion
still has some limitations, which could be overcome through using
global statistics, such as the document frequency of the query terms,
for selecting expansion terms [13].
On the one hand, query expansion without a WSD algorithm or with
a poor WSD may cause degradation in retrieval performance, which is
mainly caused by the irrelevant terms added to the query. This limitation can be alleviated by combining query expansion with local context
analysis (LCA) technique [12,13], where query expansion terms are
selected based on co-occurrence with the query terms within the topranked documents. With the combination of LCA, query expansion
terms generated from thesaurus or ontology undergo a statistical
inspection, and consequently some irrelevant expansion terms won't
be used for retrieval. On the other hand, ontology-based query expansion may miss some possible expansion terms, which are not defined
in thesaurus or ontology but have statistical dependencies with the
user's query in corpus. This limitation can be overcome by computing
statistical relevance between terms in the most relevant documents,
and then these statistical relevant terms are added to the expansion
terms for retrieval.
By combining the domain ontology and LCA technique [12,13], this
paper introduces an automatic query expansion method to improve
retrieval performance for online BIM documents.
4. Development of IFC IR Ontology
To achieve the semantic search engine for online BIM documents,
the principal work relies on two aspects: (1) how to construct a
BIM-oriented ontology for the needs of IR and (2) how to utilize IR
techniques to retrieve BIM documents based on the ontology. The
former is introduced in this section and the latter will be presented in
Section 5. This section first summarizes some basic concepts of the
domain ontology, then argues the IFC specification [11] as a semantic
foundation for our purpose, and finally introduces a method for developing the IFC IR Ontology.
In computer science and information science, an ontology is defined
as “formal, explicit specification of shared conceptualization” [35].
Ontologies can be roughly divided into two categories: general ontology
and domain ontology. The interest of general ontology is the whole
world, while domain ontology focuses on specification of particular
domain conceptualization. Although some general ontologies (e.g.,
WordNet [36]) contain a large number of general concepts, they are
not designed for domain-specific retrieval, which may lead to inaccurate description of concepts in the AEC domain. In contrast, domain
ontology is a representation of semantics in particular domain, which
17
often consists of a hierarchical description of important concepts
precisely defined in the domain, along with description of properties
of each concept [37]. Domain ontology is considered as a key element
to enhance domain-specific IR [35]. It has a variety of representation
forms, among which OWL is selected as the modeling language in this
study. With the syntax extension from RDF (Resource Description
Framework) — a lightweight representation for data and knowledge,
OWL has been proposed by W3C as the ontology language of Semantic
Web [38].
There are various methods for constructing domain ontologies, such
as TOVE, IDEF5, Skeleton, KACTUS, SEN-SUS, METHONTOLOG and
Seven-Step methods. Among them the Seven-Step method [39] is
regarded as the mature one. It was developed by the School of Medicine
in Stanford University, as well as the most widely used ontology editor
Protégé [40]. The Seven-Step method is also the only one which strictly
conforms to Gruber's five basic principles of ontology construction [35].
Following the general principles of ontology construction and the
Seven-Step method, construction of IFC IR Ontology is essentially a
process of conceptualizing and formalizing BIM knowledge from
the IFC schema. By following the ontology development process of
Seven-Step method, we give the main procedure of constructing IFC IR
Ontology as follows.
4.1. Step 1: determining the domain and scope of ontology
The first step of Seven-Step method is to define the domain and
scope of the ontology. It is an important step for minimizing the amount
of data and concepts to be analyzed, especially for the extent and
complexity of BIM-related semantics. During the ontology-design
process, it may be adjusted if necessary.
The ontology discussed in this work mainly meets the needs of information retrieval, which enables architects, engineers and other design
professionals to find useful BIM resources quickly. Therefore, it includes
specific concepts related to AEC product information belonging to the
product lifecycle phases (design, manufacturing, assembly, etc.). In
addition, since online BIM documents are generally organized by
different information providers, which often use natural languages for
describing the textual contents of documents, our ontology should
also include natural language ontologies for capturing terminologies in
documents and queries.
4.2. Step 2: selecting and reusing existing resources
The second step of Seven-Step method is to consider reusing existing
semantic resources for our particular domain and task. Several wellknown semantic resources have been developed for various AEC applications, and they have the potential to be enhanced as domain ontology
for retrieving online BIM documents [27,41]. The most notable efforts [27,41] include: ISO 12006-2, Uniclass, OmniClass, Industrial
Foundation Classes (IFC) [11], etc. Among these existing semantic
resources, structured taxonomies deserve particular attention. However, since improper taxonomies have the opposite effects leading to
confusion and difficulty to retrieve [27], a properly structured taxonomy
should be carefully selected to meet our needs.
As a relative new field, ontology research for BIM resources is still
rare, which needs more exploration. In this study, the IFC4 specification
[11] is typically selected as the backbone of IFC IR Ontology. This
paper aims to reuse rich semantic contents of newly developed IFC
specification without having to understand the complex technical
issues of IFC for information seekers. In addition, many general
ontologies such as WordNet, EuroWordNet and Cyc are already
available in electronic form, so we typically select WordNet [36] as a
common lexical database for the English language. In this paper,
WordNet will be combined with domain ontology to process terminologies in BIM documents and queries.
18
G. Gao et al. / Automation in Construction 56 (2015) 14–25
4.3. Step 3: enumerating important terms in the ontology
The third step of Seven-Step method is to enumerate important
terms in the ontology. In this work, our goal is to recognize the important terms in IFC specification to establish the ontology structure.
Initially, it is important to get a comprehensive list of terms without
worrying about overlap between concepts. Therefore, the key point
lies on how to obtain a comprehensive list of concepts from the IFC
specification.
The IFC4 specification is the latest version of IFC standard, and it
enhances the capability of previous IFC specifications in several areas
of building elements, building service elements and structural elements
and accompanying basic definitions. Currently, it contains 766 entities,
206 groups of enumeration types, 408 groups of property sets, 1691
individual properties, and lots of defined types, select types, quantity
sets, functions and rules. We confine important terms to IfcObject in
IFC specification, excluding the others. Currently, we have manually
extracted 248 entities, 140 type enumerations, and 583 enumeration
items from the IFC4 schema as the important terms of IFC IR Ontology.
However, the IFC schema was designed for computer instead of
human readability, which follows a special formalized naming convention. For example, the names of types, entities, rules and functions in IFC
start with the prefix “Ifc” and continue with the English words in Camel
Case naming convention (no underscore, first letter in word in upper
case). Therefore, they should be refined by the process of segmenting
the names of entity, removing the prefix and eliminating redundancy,
for IR purpose. This process is semi-automatically done in our work.
4.4. Step 4: defining the classes and class hierarchy
The fourth step is to define the classes and the class hierarchy. There
are three possible approaches in developing the class hierarchy [42],
including (1) the top-down approach, (2) the bottom-up approach,
and (3) a combination development process. None of these three
methods is inherently better than any of the others. Considering the
consistency of our model, we adopt the first one (i.e., the top-down
approach) in our work to avoid the duplication of works and the
inconsistency of knowledge.
Since the IFC specification has been defined as a hierarchical structure, its structure can be directly adopted in the class hierarchy of
IFC IR Ontology. On the top of the ontology, we use the EXPRESS entity
relationship directly, that is to say, we can map directly entities as
well as their subtypes and supertypes (profiting from OWL classes
inheritance). Afterwards, we treat the element's enumeration type
as its subclasses. For example, the concept “covering” has the
enumeration type (including “membrane”, “cladding”, “insulation”,
“roofing”, “modeling”, “ceiling”, “flooring”, “wrapping”, etc.) according
to the IFC schema. And, these types are connected to their superclass
(i.e., “covering”) to define the class hierarchy in the ontology.
Table 1 gives the outline of concepts in the resulting IFC IR Ontology
and Fig. 2 shows a portion of class hierarchy in the ontology. In Fig. 2, the
concept Covering inherited from the class Object has a property
PredefinedType mapped to Covering Type Enumeration. An Object is the
generalization of any semantically treated thing or process. In IFC4
specification, Object is the supertype of Actor, Control, Group, Process,
Product, and Resource, which consists of the first level of IFC IR Ontology
Table 1
The outline of concepts in the resulting IFC IR Ontology.
Taxonomies
Number of concepts
Example of concepts
Acquisition resources
Actor
Control
Group
Process
Product
Resource
2
11
12
4
211
8
Occupant
Project order
Building system
Task
Covering
Crew resource
IfcActor
IfcControl
IfcGroup
IfcProcess
IfcProduct
IfcResource
(see Table 1). In addition, IFC IR Ontology comprises a variety of
relationships, for example “PredefinedType”, “OperationType”. There are
totally 140 enumeration types related by these relationships.
4.5. Step 5: define the properties of classes
The hierarchical structure is just a frame for the concept system
which cannot reflect the varied relations between the concepts.
Therefore, we need to describe and define the properties of classes in
this step. These properties become slots attached to the classes. Slots
are the descriptors used for describing the properties of classes and
instances. We define the domain/range of the properties according to
the applicable classes and DataType in the XML specification of the
property set (Pset) of IFC 4. The OWL datatype property is used to
represent the EXPRESS simple attributes, and the OWL object property
represents named attributes. As an example of properties in IFC IR
Ontology, the properties of class “door” are defined with “Acoustic
Rating”, “Thermal Transmittance”, “Status”, “Self Closing”, “Smoke
Stop”, “Durability Rating”, “Has Drive”, “Is External”, “Glazing Area
Fraction”, “Hydrothermal Rating”, “Infiltration”, “Fire Exit”, “Security
Rating”, “Handicap Accessible”, “Reference” and “Fire Rating”.
4.6. Step 6: defining the facets of the slots
The slots of classes may have different facets to describe the value
type, allowed values, the number of the values (cardinality), and other
features of the values the slot can take. The conversion of EXPRESS
simple types (String, Integer, Real, Binary, Boolean, Logical) is direct,
as they have equivalents in OWL (XSD types). For instance, String type
is directly mapped into xsd:string type. And, the value of “Fire Rating”
in IFC (in a common property set of “door”) is IfcLabel, we directly
map it to xsd:string type. As a result, “Fire Rating” becomes a slot of
class “door”, whose type is string in OWL.
Besides mapping the value type of the slots from EXPRESS to OWL,
we also represent the allowed value types with OWL rules, and
the allowed number of the values with OWL cardinality restriction according to the EXPRESS attribute's optional flag.
4.7. Step 7: creating the instances
The last step of Seven-Step method is to create individual instances
of classes in the hierarchy. However, since our goal in this study is
information retrieval of unstructured BIM documents with the help of
IFC IR Ontology, we don't have to use the instances of IFC files.
Consequently, converting an IFC file to its ontology instance is not
necessary for our current application. Of course, it is also interesting to
explore information retrieval of IFC instance files, which is the future
work in BIMSeek development.
5. Automatic query expansion process
With reference to IFC IR Ontology, this section presents an automatic query expansion method for retrieving online BIM documents.
Starting with query words as the input in the search engine, our
algorithm mainly consists of four steps: generating the candidate
concepts, expanding the candidate concepts, query pruning and
keyword matching. The main procedure of our algorithm is shown
in Fig. 3. The following will introduce each step in detail.
5.1. Step 1: generate the candidate concepts
The users may not always formulate their search queries using
the standardized concepts in IFC specification, so the original query
terms are not suitable for direct searching. To execute the semantic
search, we should identify the “meaning” of every term in the user's
query, so that a matching concept in the IFC IR Ontology can be found.
G. Gao et al. / Automation in Construction 56 (2015) 14–25
19
Fig. 2. Part of the classes and class hierarchy in IFC IR Ontology.
This is a necessary step to semantic retrieval, through which semantics
of the terms appearing in the user's query will be automatically replaced
by the standardized concepts in the ontology.
The first step of our algorithm preprocesses the simple noun phrases
in the user's query, and generates the candidate concepts corresponding
to IFC IR Ontology. In our implementation, we first use WordNet to find
the synonyms for each query term, and then generate the candidate
concepts of the query term by matching the domain ontology. WordNet
[36] is a large lexical database of English, in which nouns, verbs,
adjectives and adverbs are grouped into sets of cognitive synonyms
(synsets). With reference to a generic lexicon in WordNet, a list of
candidate concepts can be produced for further retrieval. For instance,
the terms “cover”, “screening”, “masking”, “coating” and “application”
are the synonyms of “covering” in WordNet, while only “covering” is
the standardized concept in IFC IR ontology. When some synonyms
words of “covering” appear to the user's query, the “covering” will be
located as a candidate concept.
However, WordNet, as the generic lexicon rather than domain
ontology, often contains a large number of concepts that are not
relevant to the target domain. As a result, these irrelevant concepts
may lead to an inaccurate estimation of concept specificity and information retrieval. For example, “application” should not be regarded as the
candidate concept of query “covering” in BIM-specific domain. To
address this issue, a context-based method for concept identification
can reduce the incorrect mapping, which will be illustrated in
Section 5.3.
extent of the neighborhood is governed by the relatedness function, as
will be shown in Eq. (1). Especially, each expanded concept in the
neighborhood is associated with a relatedness value according to its
distance to the candidate concept. There are various relationships
(including subclass, superclass, type enumeration etc.) between
concepts in IFC specification. Therefore, the concepts, which have the
above IFC relationships with each candidate concept, will be added
to the new query terms, i.e., expanded concepts. Here, a semantic relatedness function between the candidate concept and its expanded concept is addressed for measuring the expansion range. The expansion
process will be terminated when the value of semantic relatedness is
less than a predefined threshold.
There are several measures of semantic relatedness, such as Leacock–
Chodorow measure [43], Jiang–Conrath measure [44], Lin measure [45],
Resnik measure [46] and Wu–Palmer measure [47]. The measures of
Lin, Resnik and Jiang–Conrath are based on the notion of information
content. In contrast, the Leacock–Chodorow measure is based on pathlength, which was found more effective than information content-based
relatedness measures [46]. In this section, we choose the Leacock–
Chodorow measure [43] for building the semantic relatedness function,
where the semantic relatedness between two concepts is estimated
based on the conceptual links (i.e., the distance) between these concepts
with reference to IFC IR Ontology. The smaller the distance between two
concepts, the higher the semantic relatedness between the two concepts
will be. Therefore, the semantic relatedness is inversely proportional
to the semantic distance in the ontology. Our semantic relatedness function
is defined by
5.2. Step 2: expand the candidate concepts
Based on WordNet, Step 1 only generates some candidate concepts
for the user's query in a level of word-concept matching. In information
retrieval, the user-entered query is usually simple and short, and it is
hard for a search engine to completely express the user's information
needs. In order to represent the information seeker's needs in BIMspecific domain, the candidate concepts can be further expanded by
utilizing the relationships between concepts in IFC IR Ontology. The
expanded concepts will be added to the query to improve the accuracy
and coverage in retrieval.
In the second step of our algorithm, a concept expansion algorithm is
proposed based on the relationships between concepts in IFC IR
Ontology. The key task lies on how to compute the relatedness between
two concepts in the ontology. In our implementation, each candidate
concept is expanded through its neighborhood over the ontology to
yield more concepts close to the candidate concept semantically. The
RelIFC ðc; ci Þ ¼ − log
lenðc; ci Þ
;
2Depth
ð1Þ
where c is a query concept in IFC IR Ontology, ci is the i-th concept related
to c in its neighborhood, Depth is the maximal tree depth of the ontology.
Especially, len(c, ci) is the shortest path between c and ci in IFC IR
Ontology, which is calculated as the edge number on the shortest path
(in practice, plus 1 to avoid that the value is zero). In Eq. (1), RelIFC(c, ci)
is the semantic relatedness function which takes into account the shortest
path len(c, ci) between two concepts c and ci defined in IFC IR Ontology.
The factor of 2 in the denominator is used to represent the possible
maximum length between two concepts. The distance between two
concepts reaches the maximum length when they are both leaf nodes
and the only common ancestor of them is the root node. This definition
guarantees that in any circumstances, the relevance value between two
concepts won't be less than 0.
20
G. Gao et al. / Automation in Construction 56 (2015) 14–25
Fig. 3. The main procedure of our algorithm.
As an example in Fig. 2, the concept “covering” can be expanded to
some subclasses (“membrane”, “cladding”, “insulation”, “roofing”,
“modeling”, “ceiling”, “flooring”, “wrapping”, etc.) in the 1-level
hyponym expansion mode. Considering the concepts “covering” and its
subclass “membrane”, the Depth is equal to 6 and len(⋅) is equal to 2.
Accordingly, the IFC relatedness between “covering” and “membrane” is:
5.3. Step 3: prune the irrelevant concepts
′′
′′
RelIFC ‘‘covering ; ‘‘membrane ¼ −log
1 The “incorrect mapping” problem. As discussed in Section 5.1,
WordNet, as a generic lexicon rather than domain ontology, often
contains a large number of synonyms that are irrelevant to BIMspecific domain. As a result, these irrelevant concepts may lead to an
inaccurate estimation of concept specificity and information retrieval.
2 The “overexpansion” problem. The expansion of concepts is based on
the relationships defined in the IFC specification, which are usually
2
¼ 0:778:
26
At the end of Step 2, we obtain the expanded concepts and their
semantic relatedness values, where all the expanded concepts are
saved as a set denoted by ExpansionSet.
As discussed in Step 1 and Step 2, the original query terms are
mapped into the IFC concepts, and then these concepts are expanded
to a set of related concepts to formulate the appropriate domainspecific query. However, simple concept mapping and expansion may
have some shortages as follows.
G. Gao et al. / Automation in Construction 56 (2015) 14–25
made by some professionals. These relationships are independent of
the document corpus in the domain. In practice, some expanded
concepts that are irrelevant to the original query will be harmful to
the retrieval, resulting in topic dilution of documents.
To solve the above problems, we use a statistical method to examine
the relatedness between concepts. In this step, we utilize local context
analysis (LCA) [12,13] for improving IFC-based concept expansion.
LCA acquires statistical relatedness between concepts and query terms
according to co-occurrence analysis of the concepts with the query
terms in top n passages retrieved by the query.
In our implementation, we first use a standard IR system to retrieve
the top n ranked passages in the corpus. Then these top-ranked
passages are used as the local “context” of the query terms, and the
“context” is used to analyze the co-occurrence of the related concepts
and the query terms. The concepts, expanded in Section 5.2, are
accepted according to a function of the frequencies of occurrence of
the query terms and co-occurrence concepts in the retrieved passages,
and the inverse passage frequencies for the entire collection of those
terms and concepts. Only the concepts that have strong relationship
are kept for final retrieval. Since the textual contents of online BIM
documents are usually short and each document usually has one single
topic rather than multiple topics, the passage in our study is defined as
the whole document in a webpage.
Given a user's query Q, we adopt LCA [12,13] to compute the
statistic relatedness measure between an IFC concept c and Q, as
defined by
RelLCA ðc; Q Þ ¼ ∏ ðδ þ coðc; t i ÞÞ;
t i∈Q
ð2Þ
where co(c, t i ) denotes the co-occurrence degree between the
concept c and each term ti in the query Q, which is computed using
the term frequency (tf) in the top-ranked documents and inverse
document frequency (idf) in the corpus [12,13]. δ is a small positive
constant (e.g., δ = 0.1) to avoid that the value of RelLCA is zero. Multiplication in Eq. (2) is used to emphasize co-occurrence with all
query terms.
- Against the “incorrect mapping” problem. For a concept c in the
expanded concepts ExpansionSet, generated in Section 5.2, if the average
of RelLCA values of concepts expanded from c is very small, it suggests
that these expanded concepts are statistically irrelevant to the user's
query through analyzing the document corpus. For example, as stated
in Section 5.1, the terms “cover”, “screening”, “masking”, “coating” and
“application” are initially mapped to the IFC concept “covering”.
However, the retrieval documents using query term “application”
barely include any other related concepts of “covering”. Hence the
mapping from term “application” to concept “covering” is regarded as
an incorrect mapping. Then all the concepts expanded from “covering”
in ExpansionSet, including “membrane”, “wrapping”, etc., should be
removed.
- Against the “overexpansion” problem. The concepts in
ExpansionSet, which have the small values of RelLCA (e.g., less than
a predefined threshold 0.3), are removed from the ExpansionSet.
This means that the expanded concepts with small RelLCA values
are not quite relevant to the original query. After pruning the irrelevant concepts, the remaining part in ExpansionSet is added to the
final query terms. For example, the mapping from the query term
“masking” to IFC concept “covering” can be accepted according to
the above principle. However, some of the expanded concepts, like
“skirting board” has the relatively small RelLCA values, hence are
abandoned from ExpansionSet in our algorithm.
After pruning all irrelevant concepts from the ExpansionSet, the final
relatedness weight of each concept in ExpansionSet can be calculated by
combining Eq. (1) and Eq. (2). Given a query concept c and its expanded
21
ci in ExpansionSet, the relatedness weight between ci and the user's
query Q is defined by
α⋅RelIFC ðci ; cÞ⋅RelLCA ci;Q
;
ω ci;Q ¼
α⋅RelIFC ðci ; cÞ þ RelLCA ci;Q
ð3Þ
where α is the importance factor (we typically set α = 1.0 in our
implementation).
5.4. Step 4: rank the documents of BIM resources
After the user's query is expanded by IFC IR Ontology and LCA,
document ranking can be implemented based on each document's
score against the query. The ranking process uses the vector space
model (VSM) [14] to determine how relevant a given document is to
the user's query. In the process, a Boolean model is first used to narrow
down the documents that need to be scored based on the use of Boolean
logic in query specification. Then, these resultant documents are ranked
based on their matching scores between the query vector and their
document vectors using the VSM. In the VSM implemented in this
research, the expanded concepts are associated with the weights
computed in Eq. (3). The score between the user's query Q and a document D in corpus is denoted by
2
3
1
1 7
2
∑6
scoreðQ; DÞ ¼ 4tf ðt; DÞ⋅idf ðt Þ ⋅ωðt; Q Þ⋅ →5
! t∈Q
VQ VD
ð4Þ
!
where t is a term in the query Q, V Q is the query vector of user's query
!
Q, V D is the document's vector of D. tf(⋅) is the term frequency in
document D, which measures how often a term appears in the document. idf(⋅) is the inverse document frequency, which measures how
often the term appears in document corpus. ω(t, Q) is the weight,
calculated by Eq. (3), measuring the importance of each query term
t in the query Q.
6. Experimental results
6.1. System overview
By combining the usability of keyword-based interface with automatic query expansion techniques, we have developed a semantic
search engine, named BIMSeek, for retrieving online BIM documents.
The presented search engine is a query-driven information retrieval
system, which automatically expands the user's query based on the
ontology and then ranks BIM document repositories by analyzing the
semantic details carried by such documents. For intuitively showing
the ontology's concepts associated with the user's query, an ontology
navigation toolbar is also developed. The toolbar provides a dynamic
navigation way, which allows the user to manually refine the query
through exploring the related concepts.
Fig. 4 shows a snapshot view of the main interface of our search
engine, which consists of four parts. First, the user inputs a specific
query on the top-left (see “Query Area”). Then, the system automatically ranks online BIM documents and displays the search results
produced by our algorithm on the bottom-left (see “Result Area”).
The ontology navigation toolbar on the top-right allows the user to
manually refine the query through clicking on the related concepts
(see “Navigation Bar”). Alternatively, one can adjust the algorithm
parameters for refining search through the control panel on the
bottom-right (see “Parameter Panel”).
In the process of building IFC IR Ontology in Section 4, we use
OntoSTEP, a Protégé EXPRESS tool [40], to semi-automatically construct
a raw IFC IR Ontology, and this raw ontology is used for further ontology
22
G. Gao et al. / Automation in Construction 56 (2015) 14–25
Fig. 4. The screenshot of our search engine. It allows a user to specify a search query using keywords. Then it returns the ranked results related to online BIM resources. The user may refine
the search through clicking on “Navigation Bar” for accessing the query-related concepts. The algorithm parameters can be alternatively adjusted through the “Parameter Panel”.
(a) Rank 1 using our method
(b) Rank 2 using our method
(d) Rank 1 using Lucene
(e) Rank 2 using Lucene
(c) Rank 3 using our method
(f) Rank 3 using Lucene
Fig. 5. Comparison of search results between our method and traditional keyword-based search (Lucene [48]). For the same test query “railing”, the webpages corresponding to top 3 query
results (from left to right) are typically selected for displaying, where the top row shows our ranking results and the bottom row shows the ones using Lucene system.
G. Gao et al. / Automation in Construction 56 (2015) 14–25
6.2. Evaluation
In the experiments, the comparison between our method and
keyword-based search has been conducted. To measure the performance of the keyword-based search, Lucene [48] with the default
function was used. Lucene is frequently used as a baseline keywordbased system in IR. The Lucene [48] scoring uses a combination of the
VSM and the Boolean model to determine how relevant a given document is to a user's query.
Fig. 5 shows an example of search results with the test query
“railing” applied to the benchmark collection of online BIM documents.
We typically display the webpages corresponding to top three query
results (from left to right), where the top row is our ranking results
and the bottom row is the ones using the keyword-based search. For
the keyword-based search, only the webpages that contain the
keywords “railing” or “rail” are returned. However, some top search
results, such as the webpage “Heavy-Duty Screen Doors” in Fig. 4 that
just has a rail material description, don't sufficiently reflect user's search
intent (“railing”). In contrast, using our method, some ontology's
concepts (e.g., “balustrade”, “handrail” and “guardrail”) associated
with the input query “railing” are expanded as new query terms.
Consequently, the most relevant webpages, which are not only similar
to the input query word (i.e., “railing”) but also semantically satisfying
the expanded concepts (e.g., “guardrail” or “balustrade”), tend to be
ranked at the top in our system. For example, rank 1 in Fig. 4 is a
webpage about “aluminum Wall Mounted Handrails”, rank 2 in Fig. 4
is the one about “stainless steel handrail”, and rank 3 in Fig. 4 is the
one about “an indoor/outdoor LED-based handrail”. In this way, BIM
documents, which do not only hit the query keywords but also semantically satisfy an information seeker's needs, can be found. The example
suggests that our retrieval method has an obvious advantage over the
keyword-based search in BIM-specific domain.
In order to evaluate the effectiveness of the proposed method, we
run an experiment of information retrieval on a document collection.
Currently, the document collection contains a total number of 15,176
BIM documents acquired from Autodesk Seek [2]. Autodesk Seek provides three industry-standard classifications (including MasterFormat,
OmniClass and UniFormat) for browsing BIM model catalogue. In this
test, we typically select OmniClass as the baseline classification of BIM
documents for our retrieval evaluation. The OmniClass number of each
BIM document is obtained through crawling the categories from the
website, and the OmniClass number is used for “ground truth” for
each test query. For instance, the OmniClass number “23.35.00.00” has
its name “Covering, Cladding, and Finishes”. Then, we conceive of a
test query “Covering, Cladding, and Finishes”, and judge whether each
webpage in the search results belongs to the corresponding OmniClass
classification (i.e., “23.35.00.00”) or its subclass (e.g., “23.35.20.21”).
There are 30 test queries used in our test.
To measure the performance of our method, we adopt the standard
evaluation procedure from information retrieval, namely precision–recall curves, which is used for evaluating the retrieval results [49,50].
The precision–recall curves describe the relationship between precision
and recall for an information retrieval method. In the precision–recall
curve, the number of relevant documents for each query is denoted as
Relevant, the number of documents retrieved for the query is denoted
as Retrieved, and the number of relevant documents correctly retrieved
is denoted as Relevant ∩ Retrieved. Then the recall is defined as
Relevant∩Retrieved
, and the precision is defined as Relevant∩Retrieved for each
Relevant
Retrieved
experiment. It is desirable to achieve both high precision and recall,
but unfortunately this is rather difficult to achieve, especially for the
text-based retrieval problem.
We compare our method with several retrieval methods, which
include the keyword-based search (Lucene), query expansion based
on WordNet (synonyms, hyponyms and hypernyms, respectively).
Fig. 6 shows the average precision–recall curves. The results show that
our method performs better than both the keyword-based search and
WordNet-based query expansion methods.
In addition to the precision–recall curves, we also evaluate other
quantitative statistics for evaluation of retrieved results. Specifically,
we compute E-measure and F-measure [51]. F-measure (also F-score)
is a measure of a test's accuracy, and it is the harmonic mean of precision
and recall. The F-measure is defined as
Fβ¼ð1þβ2 Þ⋅
Precision⋅Recall
:
β2 ⋅Precision þ Recall
It measures the effectiveness of retrieval with respect to a user who
attaches times as much importance to recall as precision. β is a
non-negative real value denoting the times as much importance to
recall as precision. Since the recall and precision are both important in
BIM resource retrieval, we set β = 1 to let the recall and precision rate
evenly weighted.
Another score combining precision and recall is E-measure, which
means the effectiveness measure:
E ¼ 1−
α
1
:
1
1
þ ð1−α Þ
Precision
Recall
In our test, α is set as 0.5 when β is 1. Table 2 gives F-measure and
E-measure of our method and other methods. By examining the results,
we can see that our method is higher in performance compared with
other methods.
7. Conclusions and future work
Based on IFC IR Ontology and local context analysis (LCA), this paper
has introduced an automatic query expansion method for retrieving
online BIM resources. IFC IR Ontology can be used for the disambiguation of terms on online BIM documents. Ontology-based query
0.7
Keyword−based search
Our method
WordNet synonym
WordNet hyponym
WordNet hypernym
0.6
0.5
Precision
refinement. Then the built ontology is automatically converted into
OWL structured form. In the process of semantic retrieval in Section 5,
the toolkit OWLAPI, a semantic system framework, is used for handling
ontology and OWL language. The indexing and ranking are offered by
Apache Lucene, and Apache Nutch Web Crawler is used to collect online
BIM documents.
The search system developed in this work has been designed to be a
flexible tool to test different retrieval methods. The system also provides
a friendly interface for the user, in which the user can modify several
simple parameters for adjusting the search results. In this section, all
the experiments are run on a 2.80 GHz processor with 4 GB memory
under Windows 7.
23
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Recall
Fig. 6. The precision–recall curves of retrieval results using our method and other
methods.
24
G. Gao et al. / Automation in Construction 56 (2015) 14–25
References
Table 2
F-measure and E-measure of our method and others.
Method
F-measure
E-measure
Keyword-based search
WordNet synonym
WordNet hyponym
WordNet hypernym
Our method
0.180093695
0.162367191
0.189367722
0.124649668
0.24816183
0.819906305
0.875350332
0.837632809
0.810632278
0.75183817
expansion enables the addition of standardized domain terms for search
query, while LCA, through BIM document analysis, offers the broader
scope of terms related to user information needs. The experimental
results show that our method can achieve better efficiency for retrieving
BIM documents, than traditional keyword-based search and WordNetbased query expansion methods.
In the current implementation, only the inheritance relationship and
the type enumeration in the IFC ontology are used for expanding and
matching BIM documents. In our ongoing work, we are trying to use
the properties and restrictions in the ontology for searching BIM
models. One alternative way is to extract and convert each BIM model
into in RDF or OWL format, and query with SPARQL.
One of the limitations of our method is that the single IFC ontology
cannot fully cover the IR needs of BIM-specific domain because of its
multidisciplinary and multistakeholder nature. Therefore, there is a
need to combine or merge the IFC ontology with some existing AEC
ontologies such as Uniclass or OmniClass [52], so that more extensive
BIM resources can be covered. In the future, we would like to integrate
more AEC ontologies (e.g., OmniClass) in the search engine. On the
other hand, the number of expansion terms using LCA is arguable. Too
few expansion terms may have no impact, and too many will cause a
query drift. A more effective query expansion method is another future
direction of research.
The experiment described in Section 6.2 is essentially a systemoriented evaluation [21], where the benchmark collection of BIM
documents is acquired from Autodesk Seek [2] and OmniClass is used
as the baseline classification of BIM documents for retrieval evaluation.
Although such a system-oriented experiment can help evaluating the
effectiveness of particular document ranking, it cannot examine the
full operational characteristics of the proposed search engine. Therefore,
a user-centered evaluation [53] might be conducted to supplement the
system-oriented experiment by multiple humans independently.
In particular, the inter-rater agreement (or inter-rater reliability) of
document ranking can be assessed by some statistical measures such
as Fleiss' kappa [54]. The user-centered evaluation studies will be left
as part of our future work.
Acknowledgments
The authors appreciate the comments and suggestions of all
reviewers, whose comments significantly improved this paper. The authors would like to thank Dr. Pieter Pauwels at Ghent University for
discussing the problem of IFC-to-RDF conversion. The research is
supported by the National Science Foundation of China (61472202,
61272229, 61003095) and the National Technological Support
Program for the 12th-Five-Year Plan of China (2012BAJ03B07).
The fourth author is supported by the Chinese 973 Program
(2010CB328003) and the last author is supported by the Chinese 863
Program (2012AA040902).
Supplementary material
The proposed system and its demonstration can be accessed at:
http://cgcad.thss.tsinghua.edu.cn/liuyushen/ifcqe/.
[1] C. Eastman, P. Teicholz, R. Sacks, K. Liston, BIM Handbook: A Guide to Building
Information Modeling for Owners, Managers, Designers, Engineers and Contractors,
2nd Edition John Wiley and Sons, NJ, 2011.
[2] Autodesk, AutodeskSeek, http://seek.autodesk.com 2014.
[3] BIMobject, http://bimobject.com/ 2014.
[4] NBS of UK, National BIM Library, http://www.nationalbimlibrary.com/ 2014.
[5] Google, Google 3D Warehouse, http://sketchup.google.com/3dwarehouse/ 2014.
[6] SmartBIM LLC, SmartBIM, http://www.smartbim.com/ 2014.
[7] Pierced Media LC, RevitCity, http://www.revitcity.com 2014.
[8] A. Court, D. Ullman, S. Culley, A comparison between the provision of information to
engineering designers in the UK and the USA, Int. J. Inf. Manag. 18 (6) (1998)
409–425.
[9] G.J. Hahm, M.Y. Yi, J.H. Lee, H.W. Suh, A personalized query expansion approach for
engineering document retrieval, Adv. Eng. Inform. 28 (4) (2014) 344–359.
[10] Z. Li, V. Raskin, K. Ramani, Developing engineering ontology for information
retrieval, J. Comput. Inf. Sci. Eng. 8 (1) (2008) 737–745.
[11] Industry Foundation Classes (IFC), IFC4 Release Candidate 4, http://www.
buildingsmart-tech.org/ifc/IFC2x4/rc4/html/index.htm 2014.
[12] J. Xu, W.B. Croft, Query expansion using local and global document analysis,
Proceedings of ACM SIGIR'96 1996, pp. 4–11.
[13] J. Xu, W.B. Croft, Improving the effectiveness of information retrieval with local
context analysis, ACM Trans. Inf. Syst. 18 (1) (2000) 79–112.
[14] G. Salton, Developments in automatic text retrieval, Science 253 (5023) (1991)
974–980.
[15] S. Deerwester, S. Dumais, G. Furnas, T. Landauer, R. Harshman, Indexing by latent
semantic analysis, J. Soc. Inf. Sci. 41 (6) (1990) 391–407.
[16] T. Hofmann, Probabilistic latent semantic indexing, Proceedings of ACM SIGIR'99
1999, pp. 50–57.
[17] D.M. Blei, A.Y. Ng, M.I. Jordan, Latent Dirichlet allocation, J. Mach. Learn. Res. 3 (4–5)
(2003) 993–1022.
[18] J. Bhogal, A. Macfarlane, P. Smith, A review of ontology based query expansion, Inf.
Process. Manag. 43 (4) (2007) 866–886.
[19] S. Kara, O. Alan, O. Sabuncu, S. Akpinar, N. Cicekli, F. Alpaslan, An ontology-based
retrieval system using semantic indexing, Inf. Syst. 37 (4) (2012) 294–305.
[20] O. Egozi, S. Markovitch, E. Gabrilovich, Concept-based information retrieval using
explicit semantic analysis, ACM Trans. Inf. Syst. 29 (2) (2011) (Article 8).
[21] X. Yan, R. Lau, D. Song, X. Li, J. Ma, Toward a semantic granularity model for
domain-specific information retrieval, ACM Trans. Inf. Syst. 29 (3) (2011) (Article
15).
[22] P. Demian, P. Balatsoukas, Information retrieval from civil engineering repositories: importance of context and granularity, J. Comput. Civ. Eng. 26 (6)
(2012) 727–740.
[23] K. Lin, S. Hsieh, H. Tserng, K. Chou, H. Lin, C. Huang, K. Tzeng, Enabling the creation of
domain-specific reference collections to support text-based information retrieval
experiments in the architecture, engineering and construction industries, Adv.
Eng. Inform. 22 (3) (2008) 350–361.
[24] K. Lin, L. Soibelman, Incorporating domain knowledge and information retrieval
techniques to develop an architectural/engineering/construction online product
search engine, J. Comput. Civ. Eng. 23 (4) (2009) 201–210.
[25] H.-T. Lin, N.-W. Chi, S.-H. Hsieh, A concept-based information retrieval approach for
engineering domain-specific technical documents, Adv. Eng. Inform. 26 (2) (2012)
349–360.
[26] A. Weissman, M. Petrov, S. Gupta, A computational framework for authoring and
searching product design specifications, Adv. Eng. Inform. 25 (3) (2011) 516–534.
[27] Y. Rezgui, Ontology-centered knowledge management using information retrieval
techniques, J. Comput. Civ. Eng. 20 (4) (2006) 261–270.
[28] L. Zhang, R. Issa, Ontology based partial building information model extraction, J.
Comput. Civ. Eng. 27 (6) (2013) 576–584.
[29] J. Beetz, J. Leeuwenand, B. Vries, IfcOWL: a case of transforming EXPRESS schemas
into ontologies, Artip. Intell. Eng. Des. Anal. Manuf. (AI EDAM) 23 (1) (2009)
89–101.
[30] L. Zhang, R. Issa, Development of IFC-based construction industry ontology for
information retrieval from IFC models, Proceedings of the 2011 EG-ICE Workshop,
2011.
[31] P. Pauwels, D.V. Deursen, R. Verstraeten, J.D. Roo, R.D. Meyer, R.V. de Walle, J.V.
Campenhout, A semantic rule checking environment for building performance
checking, Autom. Constr. 20 (5) (2011) 506–518.
[32] M.P. Nepal, S. Staub-French, R. Pottinger, A. Webster, Querying a building
information model for construction-specific spatial information, Adv. Eng. Inform.
26 (4) (2012) 904–923.
[33] A. Borrmann, E. Rank, Specification and implementation of directional operators in a
3D spatial query language for building information models, Adv. Eng. Inform. 23 (1)
(2009) 32–44.
[34] G. Gao, Y.-S. Liu, M. Wang, X.-G. Han, BIMTag: semantic annotation of web BIM
product resources based on IFC ontology, 21st International Workshop: Intelligent
Computing in Engineering 2014 (EG-ICE), 2014.
[35] T. Gruber, A translation approach to portable ontology specifications, Knowl. Acquis.
5 (2) (1993) 199–220.
[36] G. Miller, WordNet: a lexical database for English, Commun. ACM 38 (11) (1995)
39–41.
[37] T. Lukasiewicz, U. Straccia, Managing uncertainty and vagueness in description
logics for the semantic web, J. Web Semant. 6 (4) (2008) 291–308.
[38] W3C, OWL 2 Web Ontology Language Overview, http://www.w3.org/TR/owl-features/ 2009.
G. Gao et al. / Automation in Construction 56 (2015) 14–25
[39] N.F. Noy, D.L. McGuinness, Ontology development 101: a guide to creating your first
ontology, http://protege.stanford.edu/publications/ontology_development/ontology1 01.html 2014.
[40] Protégé, An ontology and knowledge base editor, http://protege.stanford.edu/ 2014.
[41] K. Lin, L. Soibelman, Promoting transactions for A/E/C product information, Autom.
Constr. 15 (6) (2006) 746–757.
[42] M. Uschold, M. Gruninger, Ontologies: principles, methods and applications, Knowl.
Eng. Rev. 11 (2) (1996) 93–136.
[43] C. Leacock, M. Chodorow, Combining Local Context and WordNet Similarity for
Word Sense Identification, 1st EditionMIT Press, 1998 (Ch. 11).
[44] J. Jiang, D. Conrath, Semantic similarity based on corpus statistics and lexical taxonomy, Proceedings of International Conference Research on Computational Linguistics 1997, pp. 112–123.
[45] D. Lin, An information-theoretic definition of similarity, Proceeding of 15th International Conference on Machine Learning 1998, pp. 296–304.
[46] P. Resnik, Using information content to evaluate semantic similarity in a taxonomy,
International Joint Conference on Artificial Intelligence 1995 1995, pp. 448–453.
25
[47] Z. Wu, M. Palmer, Verb semantics and lexical selection, Proceedings of 32nd Annual
Meeting of the Association for Computational Linguistics 1994, pp. 448–453.
[48] Lucene, http://lucene.apache.org/ 2014.
[49] V. Raghavan, P. Bollmann, G.S. Jung, A critical investigation of recall and precision as
measures of retrieval system performance, ACM Trans. Inf. Syst. 7 (3) (1989)
205–229.
[50] C.D. Manning, P. Raghavan, H. Schütze, Introduction to Information Retrieval, Cambridge University Press, 2008.
[51] C.J.V. Rijsbergen, Information Retrieval, 2nd Edition Butterworths, London, 1979.
[52] N. El-Gohary, T. El-Diraby, Merging architectural, engineering, and construction ontologies, J. Comput. Civ. Eng. 25 (2) (2011) 109–128.
[53] T. Saracevic, Evaluation of evaluation in information retrieval, Proceedings of ACM
SIGIR'95 1995, pp. 138–146.
[54] J.L. Fleiss, Measuring nominal scale agreement among many raters, Psychol. Bull. 76
(5) (1971) 378–382.